Hybrid F-K Filtering and Deep Learning for P/S Separation in DAS VSP Data

The accurate separation of P-waves and P-S converted waves in DAS VSP data is crucial for underground imaging. Traditional methods suffer from severe artifacts, and standard polarization analysis is inapplicable to single-fiber measurements. While pure end-to-end deep learning methods are powerful, their application to actual field data is often bottlenecked by the scarcity of real-world labels. In this study, we propose FKPSN, representing a novel physics-guided task-simplification paradigm. By leveraging F-K filtering as a macroscopic kinematic prior, we fundamentally transform the end-to-end separation into a highly tractable, physics-constrained error-correction task. To explicitly address the mutual energy leakage introduced by this initial F-K separation, we tailor a dual-branch convolutional neural network equipped with a feature fusion module. This design enforces interactive cross-branch communication to simulate cross-component interference cancellation, dynamically reallocating leaked wave energy to the correct components. Extensive experiments on synthetic and field DAS VSP data demonstrate that FKPSN achieves high-fidelity separation in both time and frequency domains. Crucially, comprehensive velocity generalization tests reveal that the network possesses a robust active kinematic error-correction capability, successfully recovering true wavefields even when the initial F-K velocity parameters contain severe deviations.